AI visibility audit: how to run one, step by step

This guide is the hands-on side of AI visibility: that page explains what the metrics mean, this one shows how to collect them yourself. The examples are not made up. They are real answers we collected from ChatGPT and Gemini on 3 October 2026, and real robots.txt files we read the same day.

Our free AI visibility checker runs the answer part of this audit for ChatGPT, Gemini and Google AI Overviews, but every step below also works by hand. The example rows come from our own measurement of AI visibility tools, where neither engine named Llamat in any of 66 answers.

What is an AI visibility audit?

An AI visibility audit is a structured check of how AI assistants present your brand to people who have not asked for it by name. You ask the engines the questions your buyers ask, record which brands each answer names, recommends and cites, and compare your share with your competitors'. A complete audit also checks whether the engines' crawlers are allowed to read your website.

The term covers two different checks, and it helps to know which one you are running or buying:

• A readiness audit looks at your website: whether AI crawlers can reach it, whether the content is text they can parse, whether it is structured and current.
• An answer audit looks at the answers: whether ChatGPT, Gemini or another engine actually names you when a buyer asks.

Revenue Experts, which sells AI visibility audits, draws the same line when it describes its own free readiness audit: it “never asks, ‘Are you actually cited?’ It tells you whether the door is open, not whether anyone walked through it.” (Revenue Experts, published 30 May 2026, updated 28 July 2026, read 3 October 2026.) A site can pass every readiness check while the engines recommend three competitors and never mention it. That is why this guide starts with the answers and leaves the technical check for step 5.

You will also come across the terms AI brand visibility audit and AI search visibility audit. An AI brand visibility audit puts the weight on how the engines describe you and which competitors stand next to you. An AI search visibility audit usually means AI features inside a search engine, such as Google AI Overviews and AI Mode. The method is the same for both. Only the engines you ask change.

How to audit AI visibility in six steps

• Pick 10 to 20 prompts from real searches, without your brand name in them.
• Ask each prompt three times on each engine, every time in a new logged-out session.
• Record mention, recommendation, position and citations for every answer.
• Score the answers: rates per engine, a majority per prompt, one visibility score.
• Check that the engines' crawlers can reach your website.
• List the brands named instead of you and the sources the engines cite.

Ten prompts, asked three times on two engines, make 60 answers. A spreadsheet is enough. Our CSV template has the columns and one prompt already laid out for ChatGPT, Gemini and Perplexity, three runs each. Perplexity showed a sign-in wall instead of an answer to all 12 of our logged-out requests on 1 October 2026, so either sign in to use it or delete its rows.

Step 1: pick prompts from real searches

The prompts decide everything that follows, so take them from evidence of demand, not from a brainstorm. Start with what people type into a search engine about what you sell. Type your category into Google and note the autocomplete suggestions, then do the same for “best [category]”, “[category] for [audience]”, “[competitor] alternative” and “[category] vs”. Keep the phrases a buyer would type before they know your name, and leave your brand out: the audit measures whether the engines bring you up on their own.

Then rewrite each phrase as a question someone would ask ChatGPT, in the language of the market you sell in. Most should ask for recommendations; add a few comparisons, price questions and “how do I choose” questions. Ten prompts give a first reading, and twenty cover a category better. Write them down and keep them fixed: if the prompts change between two audits, you can no longer compare the results.

In Llamat's free report this step is automated. After you confirm your e-mail address, our research finds real search phrases about what your website offers and an AI critic keeps only the relevant ones. The kept phrases are then rewritten into up to 20 prompts in your market's language, and the report lists every one of them.

Step 2: ask each prompt three times on each engine

Open a new chat for every run, logged out or in a private window, with memory and personalization off where the engine offers them, so that earlier conversations can't steer the answer. Don't ask the same prompt three times in a row. Ask run 1 of every prompt, then run 2, then run 3. Write down the date, the engine and where you are. OpenAI says ChatGPT may use an approximate location based on your IP address to give local results (OpenAI help, read 3 October 2026).

Why three runs: one answer is one sample, and the next one can name different brands. On 3 October 2026 we asked 11 real search phrases about AI visibility tools three times on logged-out ChatGPT and three times on logged-out Gemini, then counted every tool each answer named:

• ChatGPT: of the tools it named for a phrase, 64% appeared in all three runs, 15% in two and 21% in only one.
• Gemini: 46% appeared in all three runs, 20% in two and 34% in only one.
• The two engines differed more than the runs did. For each phrase we compared the five tools each engine named most often. The two lists shared 1 tool out of 9 for “best chatgpt visibility tracker” and 4 out of 6 for “best ai visibility tools”, and 38% on average (Jaccard index: shared tools ÷ all tools on either list).

So a third of the tools Gemini named for a phrase came up in only one of three runs: a single answer would have shown or hidden them by chance. And a test on one engine would have produced a different shortlist from the other engine. Three runs is the smallest number in which no single odd answer can decide a prompt on its own. All 66 answers are in the study which AI visibility tools ChatGPT and Gemini recommend.

Step 3: record mention, recommendation, position and citations

Give every answer one row: prompt, engine, run and date, the brands it names as options in the order they first appear, and the sources it lists. Four values in that row are about you. These are the definitions our report uses, described in full in our methodology:

• Mentioned: your brand name or domain appears in the answer text.
• Recommended: the answer presents you as a choice for the reader. That includes an item in a list of suggested options, a top pick, “best for …” or “I recommend”. A passing mention doesn't count, and neither does a point of comparison, something to avoid or a source. An answer that doesn't name you never counts as recommending you.
• Position: your place among the brands the answer names as options, in order of first appearance. Named first is 1.
• Cited: one of the sources the engine lists with the answer is on your domain or one of its subdomains. An answer without listed sources counts as not citing you.

Record the other brands as well, because step 6 needs them. The first table further down shows six real rows for one prompt.

Step 4: score the answers

Score each engine on its own first. Every rate is a share of that engine's successful answers, so an answer that failed counts neither for you nor against you:

• Mention rate = answers that name you ÷ successful answers.
• Recommendation rate = answers that name and recommend you ÷ successful answers.
• Citation rate = answers whose listed sources include your domain ÷ successful answers.
• Average position = the sum of your positions ÷ the number of answers that name you.
• AI Visibility score = the average of the per-answer scores. An answer scores 0 when it doesn't name you. Otherwise it scores 0.5, plus 0.3 if it recommends you, plus 0.2 ÷ your position.

For the overall figures, average the engines with equal weight, so an engine that returned fewer answers doesn't count for less. Average position is the exception: pool every answer that names you. Then draw a heatmap of prompts by engine. A tile is ✓ when you are named in at least 2 of the 3 runs. The majority decides only the tiles: in the rates, every answer counts on its own.

Here is the arithmetic for the six real answers in the tables further down, as if Otterly.AI were the brand being audited. ChatGPT named it in all three runs, at positions 2, 3 and 2. Gemini named it in all three runs too, at positions 6, 4 and 5. Each of the six answers lists it among the tools it suggests, so each one counts as a recommendation. On ChatGPT the three answers score 0.5 + 0.3 + 0.2 ÷ 2 = 0.90, 0.5 + 0.3 + 0.2 ÷ 3 = 0.87 and 0.5 + 0.3 + 0.2 ÷ 2 = 0.90: (0.90 + 0.87 + 0.90) ÷ 3 = 89%. On Gemini they score 0.5 + 0.3 + 0.2 ÷ 6 = 0.83, 0.5 + 0.3 + 0.2 ÷ 4 = 0.85 and 0.5 + 0.3 + 0.2 ÷ 5 = 0.84: (0.83 + 0.85 + 0.84) ÷ 3 = 84%. Both tiles are ✓.

The example shows what the score rewards. Being named sixth instead of second cost Otterly.AI only 0.07 per answer, but not being named costs the whole answer. On the same prompt, Ahrefs was named in all three ChatGPT answers and in none of the three Gemini answers. Read the score next to the mention and recommendation rates, never instead of them. One prompt is one row of the heatmap; a real audit computes the rates over all its prompts.

Step 5: check that AI crawlers can reach your website

Open yourdomain.com/robots.txt. Look for groups that name the crawlers behind the engines you audit, and look at the catch-all group, User-agent: *, which applies to every crawler the file doesn't name. Here is what each crawler controls, according to its operator's documentation, read on 3 October 2026:

• OAI-SearchBot: OpenAI uses it to show websites in ChatGPT's search features. Sites that block it are not shown in ChatGPT search answers, although they can still appear as navigational links. OpenAI says changes take about 24 hours.
• GPTBot: crawls content that may be used to train OpenAI's models. Blocking it is a training opt-out, and OpenAI says it is independent of OAI-SearchBot.
• ChatGPT-User: visits a page when a ChatGPT user's question calls for it. OpenAI says robots.txt rules may not apply to it.
• Googlebot: Google names its robots.txt rules as the control for AI features in Search, which include AI Overviews and AI Mode. A page must be indexed and eligible to show with a snippet, and Google says there are no additional technical requirements.
• Google-Extended: a robots.txt token, not a separate crawler. Google says it controls whether crawled content may be used to train future Gemini models and for grounding in the Gemini apps, and that it does not affect inclusion or ranking in Google Search.

Two more checks for Google. In Search Console, the Search generative AI control under Settings is set to include your site by default; if someone excluded the site, its links and content stop appearing in AI Overviews and AI Mode. Google says the control was rolled out to all websites as of 31 August 2026. And bot protection at your CDN or host can stop crawlers that robots.txt lets in. Google's page on AI features lists “ensuring that crawling is allowed in robots.txt, and by any CDN or hosting infrastructure” among the SEO basics that still apply.

How common these blocks are: on 3 October 2026 we read the robots.txt files of the 10,000 most popular websites in the Tranco ranking and could tell the rules for 6,528 of them. 11.7% fully block GPTBot and 5.1% fully block OAI-SearchBot. 149 of those 333 OAI-SearchBot blocks come only from a catch-all rule that never names the crawler, so read the * group, not only the named ones. 484 of the 585 sites that block Google-Extended still let Googlebot in, and only 1.7% block Googlebot. The full study is which websites block AI crawlers, and the fixes are in our guide to LLM SEO.

A crawler check tells you whether the engines may read your site. It doesn't tell you whether they name you; steps 1 to 4 do that.

Step 6: find the competitor gap

Go back to the brands you recorded in step 3. For each one, count the answers that name it on each engine and divide by that engine's successful answers, the same way as for yourself, then sort the list. The brands at the top are your competitors in AI answers, whoever you compete with in search.

Keep the engines apart, because their lists differ. In our 3 October measurement, Scrunch was named in 18 of 33 ChatGPT answers but in 2 of 33 Gemini answers. SE Ranking / SE Visible went the other way: 7 of 33 on ChatGPT, 16 of 33 on Gemini. Then look at three things:

• Prompts where competitors are named and you are not. These are the empty rows of your heatmap and the first gap to close.
• What the recommended brands are praised for: price, a feature, fit for a use case. Engines can only repeat reasons someone has written down, so check whether your own pages and the reviews about you state the same facts.
• Which sources the engines cite for the topic. They lean on different ones: in the same measurement, ChatGPT cited dataforaisearch.com in 14 of 33 answers and ahrefs.com in 10, while Gemini cited zapier.com in 19 and reddit.com in 14. Being covered on the pages each engine reads is part of the fix.

End the audit with a short list: the prompts to win, the pages to write or fix and the third-party pages to get onto. After the changes, ask the same prompts again in the same way.

How often to run an AI visibility audit

There is no fixed rule. Run it again after a change you want to evaluate, leaving time for engines that search the web to find the new pages. To catch shifts you didn't cause, also run it at a steady interval, such as monthly. Answers move even when nothing on your side changes. Between 1 and 3 October 2026, logged-out ChatGPT kept the same five most-named tools for four phrases we asked on both days, but the number of answers naming Promptwatch fell from 6 of 12 to 2 of 12. Each time, reuse the same prompts, engines and market.

If you need the numbers every week, a tracker costs less effort than repeated audits by hand: repeating the same checks on a schedule is AI brand monitoring. Our AI visibility tool evaluation checklist lists what to ask before you pay for one, and our comparison of AI visibility tools covers 22 of them.

The free report: the answer audit, done for you

Llamat's free AI visibility checker automates steps 1 to 4 for ChatGPT, Gemini and Google AI Overviews and adds the competitor list from step 6. You enter your website and e-mail address. Nothing happens until you confirm the address from the link we send, which is valid for 48 hours. Then our research turns real search phrases about what your site offers into up to 20 prompts, and each prompt is asked three times on ChatGPT and three times on Gemini, every time in a new logged-out session, and searched once on Google for its AI Overview: up to 140 answers. The results usually arrive by e-mail within 30 minutes of your confirmation. The prompt research takes two to four minutes; most of the rest is the Google searches for AI Overviews, which run one at a time because Google allows only about one search a minute.

Mentions and citations are matched by code in the answer text and its listed sources. The list of competitors and whether an answer recommends you come from our own language model, which runs on our own server and lists at most 10 brands per answer. ChatGPT answers come from the logged-out web app through our own proxy farm with Czech IP addresses: ChatGPT has no country setting, and the answer language follows the prompt. Gemini answers come from the same Czech IP addresses, even though we pass your market's country. The Google searches for AI Overviews leave from the same addresses, with your market's country and language set, and a search where Google shows no AI Overview counts as an answer that doesn't name you. Every formula is in our methodology.

What it doesn't do: the crawler check from step 5. It is not a daily tracker either; its free dashboard repeats the same prompts once a week. Google AI Mode, Perplexity, Claude and Microsoft Copilot are available as paid add-ons on request.